Google Unveils Gemma 3: Powerful Open AI Models for Single Devices
Google introduces Gemma 3, a highly capable family of open AI models built on Gemini 2.0 technology that runs efficiently on a single GPU or TPU.
Google releases Gemma 3, a new lineup of lightweight, state-of-the-art open models built on the same research behind Gemini 2.0. This release celebrates over 100 million downloads of the Gemma family and offers models in 1B, 4B, 12B, and 27B sizes to fit various hardware requirements. Developers run these advanced models directly on devices like phones, laptops, and workstations.
Despite their compact size, Gemma 3 models outperform much larger competitors like Llama3-405B and DeepSeek-V3 in preliminary human preference evaluations. The models support over 140 languages and feature a 128k-token context window for processing long documents. They also include advanced text and visual reasoning capabilities alongside function calling to handle complex application tasks.
To maximize efficiency, Google provides quantized versions of Gemma 3 that reduce computational requirements while speeding up performance. The release also introduces ShieldGemma 2 to promote responsible AI development. These tools give developers the necessary resources to build fast, capable AI applications wherever users need them.